Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Is Last Minute Bidding Bad?
Joseph Tao-yi Wang∗
Humanities and Social Sciences, Caltech
November 17, 2006
Abstract
We demonstrate how last minute bidding on internet auctions is not compatible
with a private value setting with a proxy bidding system, but can be rationalized by
adding another identical auction. In the repeated eBay auction model, the last minute
bidding equilibrium, in which bidders only bid at the last minute in the first auction,
is the unique symmetric equilibrium considering monotonic, undominated strategies.
Nonetheless, though high valuation bidders efficiently win the items, under strict af-
filiation, expected revenue is lower in the first auction than in the last, and sellers can
increase their revenue by running a joint multi-unit auction. (JEL C73, D44, L81)
Keywords: Sequential auctions, Sniping, eBay auctions
∗(Email: [email protected]) This paper is derived from the first chapter of my dissertation. I thank
my dissertation chair John G. Riley for his guidance, patience and support. I would also like to thank Robert
Porter, who went through an earlier version thoroughly to provide detail comments. All remaining errors are
my own.
1
Pluralitas non est ponenda sine necessitate.
Plurality shouldn’t be posited without necessity.
—William of Occam (1285–1349)
During the past ten years, online auction houses such as eBay help millions of
buyers and sellers meet and trade with each other. One interesting phenomenon of
eBay auctions that has not been fully understood is “sniping,” or last minute bidding,
meaning that most buyers submit their bids very late during these online auctions.1 As
Patrick Bajari and Ali Hortacsu (2003) note, “more than 50 percent of final bids are
submitted after 90 percent of the auction duration has passed.”
Intuitively, if there were overlapping auctions, it would be natural to see people
bid on the auctions that end first, and then bid on the next auction if they did not win
the first one. This would result in bids coming only after all previous auctions are
gone, and hence, “late.” However, sniping refers not to “late” bidding, but attempts to
bid almost exactly on the last second. In fact, attempting to bid at the last second is a
common practice on all kinds of items. There are even various specialized softwares
online that place last minute bids for you.2
The standard auction theory with private value (PV) has little to say about last
minute bidding. Since auction houses like eBay have proxy bidding systems, buyers
submit a proxy bid as their maximum and the eBay robot matches new incoming bids
up to this maximum. Hence, setting a proxy bid of one’s valuation as soon as he or she
sees the auction would be a dominant strategy.3 Nevertheless, last minute bidding is
commonly observed in all categories of items, common value and private value alike.
Moreover, most bidders rationalize last minute bidding saying it can “avoid trig-
gering a price war between each other.” On the contrary, standard auction theory tells
1See Alvin E. Roth and Axel Ockenfels (2002) for a full description of this phenomenon.2For example, a keyword search of “snipe” on google will give you sponsor links to www.ezsniper.com,
www.bidnapper.com, www.auctionsniper.com, www.powersnipe.com, etc.3This is why Bajari and Hortacsu (2003) think it is crucial private value models (such as Ockenfels and
Roth (2006)) require additional assumptions such as lost last minute bids.
2
us that the timing of the proxy bids does not matter as long as everyone bids their val-
uation, and hence, such common sense rationale should not hold.4 Interesting enough,
the common sense rationale assumes the possibility of a price war if bidders bid early,
hinting that bidders would “revise” their valuation and bid aggressively if someone
else bids early, and would not do so if they bid on the last minute. This raises the
question: Is this simply a misperception, or is there something that is not captured in
the standard model?
On the other hand, it would not be surprising to see similar or identical items sold
repeatedly or sequentially5 for auction houses like eBay where thousands of items are
auctioned off daily. If there are two identical auctions conducted in a row, bidders
need not want to bid up to their true valuation in the first auction if they expect to
win the second one at a lower price. Hence, their maximum willingness to bid for
the current auction would not equal to their true valuation. In fact, it would be the
expected winning price of the next auction, which is the expectation of the highest
loser’s valuation in the next auction, conditioning on being the tied winner of the
current one. Since this expectation is correlated with other bidders’ valuations, it
creates a “common value component” in the private value setting, and brings in all of
the information issues we might face in a common value environment.6
The idea of identical auctions conducted simultaneously or repeatedly is not new.
Paul R. Milgrom and Robert J. Weber (2000) consider three different types of simul-
taneous auctions and four sequential (or repeated) ones. Under the affiliated private
value (APV) setting, they proved that the winning price path of sequential first price
4In fact, with a minimum increment, bidders would actually prefer to bid as early as possible.5A “repeated auction” is a sequence of identical single-unit auctions conducted repeatedly, and is what
Milgrom and Weber called “sequential auctions.” Here, we use the terms “repeated auction” and “sequential
auction” interchangeably. However, note that the term “sequential auction” might also refer to other types of
auction designs, such jump bidding followed with a button auction in Christopher Avery (1998).6In fact, this is why repeated auctions act similar to common value auctions, and as shown below, induce
last minute bidding behavior.
3
auctions with price revelation follow a martingale.7 Robert J. Weber (1983) character-
ized the equilibria of various sequential auctions under the independent private value
(IPV) assumption and proved the revenue equivalence theorem behind these auctions.
Orley Ashenfelter (1989) describes repeated wine auctions which produced decreasing
prices within seconds apart, and R. Preston McAfee and Daniel Vincent (1993) show
how increasing risk aversion can create this so called, “decreasing price anomaly.”
However, most literature on empirical auctions abstract away from repetition and stick
to the single auction model. Although repetition may not be important in traditional
auctions, it should play an important role in internet auctions since an easy search tool
could locate hundreds or even thousands of similar auction conducted repeatedly.
In this paper we follow the repeated auctions literature and construct a repeated
auction model to explain last minute bidding in eBay auctions. In repeated eBay auc-
tions with affiliated private value (APV), bidders only bid at the last minute and shade
their bids in the early auction. However, the increasing equilibrium results in the high-
est bidders winning the auctions, and hence, the ending rule of eBay does not hurt
efficiency. On the other hand, with “strict” affiliation, the expected revenue of the first
auction is less or equal to that of the last,8 in which equality is maintained when val-
uations are, say, independent. Moreover, expected revenue of the last auction would
equal to the unit price in a multi-unit Vickrey auction (selling the two items simulta-
neously). Therefore, the seller of the first auction would be better off to conduct her
auction later or together with the other seller when there is “strict” affiliation. How-
ever, such revenue ranking is eliminated when valuations are independent, and revenue
equivalence, as shown in Weber (1983) is restored.
In contrast, we show that without repetition, neither proxy bidding nor the fixed
7They also showed that the repeated button auction would lead to full information revelation, and exact
same winning prices. Unfortunately, their conjectures about other sequential auctions could not be verified.
See their own discussion in Milgrom and Weber (2000).8This revenue difference is closely related to the declining price anomaly described in Ashenfelter (1989),
and McAfee and Vincent (1993) provide some possible explanation.
4
ending time per se, could lead to last minute bidding behavior we actually observe.
Hence, it is indeed the case that “plurality is introduced only when necessary.”9
There are other various alternative explanations of last minute bidding. In the com-
mon value setting, Axel Ockenfels and Alvin E. Roth (2006) showed an example in
antique auctions where amateurs would bid after experts who have more information,
and thus inducing the experts to bid at the last minute to avoid facing a bidding war
against the amateurs. Julia Schindler (2003) elaborates on this idea of withholding
information and shows how different interdependent value structures can induce last
minute bidding. These are all under the standard auction framework.
In the private value setting, however, additional assumptions are needed to induce
last minute bidding. For example, Ockenfels and Roth (2006) add the assumption that
the ending period has a positive probability p that a last minute bid is not successfully
transmitted. With the probability p as common knowledge, they showed another ex-
ample of an equilibrium where both bidders jump in at the last minute bidding their
valuations and hoping that only their own bids are successfully transmitted and the
other’s bid is lost. Since there is a positive probability of losing some bids, sellers’
expected revenue, and hence, eBay’s service charges, are lowered. Moreover, overall
efficiency is not always obtained since there is a positive probability that the highest
bid is not transmitted. In this case, one might wonder why auction houses would not
try to reduce or even eliminate the possibility of lost bids since it is profitable for them
to do so.10
On the other hand, Eric Rasmusen (2006) assumes that uninformed bidders ini-
9This principle, in its more modern form, is usually referred as the Occam’s razor, though there is no
direct evidence that William of Occam himself ever wrote it down.10One possible explanation is the auction house wants to keep the bidders from leaving the auction house.
However, it is not clear why bidders would leave because their bids are not less likely to be lost. See Dan
Ariely, Axel Ockenfels and Alvin E. Roth (2005) for other reasons they suggest. Also note that this “collusive
bidding” equilibrium supported by the non-transmission of the last minute bid is just one of the (possibly
many) equilibria.
5
tially do not know their private valuation and has to incur a cost to find out. When
facing an informed bidder who knows his own value, the uninformed bidder might be
able to “save” the cost of finding out if the informed bidder has a high valuation. This
timing of finding out one’s value induces last minute bidding behavior. Although the
discovery of one’s value shares similar intuition as in the last minute revision in our
paper, the setting is rather specific and special.11 In contrast, the repetition assumption
is rather natural and general.
The rest of the paper is consisted as following. Section I describes a simple ex-
ample to illustrate the main results. Section II describes the repeated eBay auction
format, while section III characterizes the last minute bidding equilibrium, and hence,
provides an explanation for last minute bidding. Section IV discusses other auction
formats showing that we cannot easily explain last minute bidding without repeated
auctions, and section V concludes.
I. An Example: v1, v2, v3 ∼ iid uniform [0, 1]
Before we go to the general setting, we illustrate our main results with a simple exam-
ple. Consider the case where there are three bidders with unit demand bidding on two
identical items sold sequentially by two sellers. Bidders have valuations v1, v2 and v3,
independently drawn from the distribution uniform [0, 1], while sellers do not value
the items at all.
The first item is auctioned off through a second price auction where there are mul-
tiple rounds of bid submission. After each round of bid submission, the high bidder
is revealed, as well as the current price, which is the second highest bid (up to now).
Then a next round of bid submission is allowed, save new bids have to be strictly
11Here, it is crucial that the uninformed bidder faces an informed one. If both bidders are uncertain
about their values, to get last minute bidding behavior, this model requires bidders only get imperfect signals
about their private values, which is actually not private values, but (independent) “affiliated value” in Paul R.
Milgrom and Robert J. Weber (1982) general model. See Eric Rasmusen (2005).
6
higher than the current price. At the end of the last round, the bidder who submits the
highest bid wins the item and pays the then current price. After the first item is sold,
the procedure is repeated for the second one as well.
In the second auction, only two bidders are left. In that case, bidders would bid
their valuation during a certain round of bidding since not doing so is a dominated
strategy, and hence, the winner would pay a price equal to the other bidder’s valua-
tion.12 Considering only increasing equilibrium, the bidder with the highest valuation
V(1) would win the first item, leaving V(2) and V(3) to fight over the remaining item.
Hence, the price of the second item would be V(3).
Now we consider the first auction. Without loss of generality, consider vi = v1.
Without further information, bidder v1 would not be willing to pay a price higher than
what she expects to pay in the second auction (when she is tied to win), which is
E[V(3)|V(1) = v1, V(2) = v1] = E[V(3)|V(2) = v1] (by independence)
or the expectation of the third highest valuation conditional on being the (tied) winner
of the first auction.
We now calculate E[V(3)|V(2) = v1] by first obtaining g(v(3)|V(2) = v). Since
G(y) = Pr(V(3) < y|V(2) = v)
=2
∫ y0 dv2
∫ 1v dv3
2∫ v0 dv2
∫ 1v dv3
=y
v
we have
g(y|V(2) = v) =1v
for 0 ≤ y ≤ v.
Hence,
E[V(3)|V(2) = v1] =∫ v
0
y
v1dy =
v1
2
On the other hand, seeing bidder 2 bid b in earlier rounds, and (in equilibrium)
correctly inferring that she has V2 ≥ b for b yields a revision of G(y|v). If b ≤ v, the
12They would even bid in the first round if there is, say, a minimum increment that might prevent them
from bidding in later rounds.
7
revision is:
G(y|v) = Pr(V(3) < y|V(2) = v1, V2 ≥ b)
=
∫ yb dv2
∫ 1v dv3 +
∫ y0 dv3
∫ 1v dv2∫ v
b dv2
∫ 1v dv3 +
∫ v0 dv3
∫ 1v dv2
=
2y−b2v−b , if b ≤ y ≤ v,
y2v−b , if 0 ≤ y ≤ b.
Hence, we have
g(y|v) =
22v−b , if b ≤ y ≤ v,
y2v−b , if 0 ≤ y ≤ b ≤ v.
Therefore, for b ≤ v1,13
E[V(3)|V(2) = v1, V2 ≥ b]
=∫ b
0
y
2v1 − bdy +
∫ v1
b
2y
2v1 − bdy
=v1 − b2
2v1
2− bv1
≥ v1
2.
In other words, when a bidder bids early, other bidders who have a valuation higher
than the early bid would have a new G(y|·) first order stochastically dominating the
old one, and hence, have a higher E(y|·). Therefore, bidding in earlier rounds would
raise high valuation opponents’ conditional expectation of next auction’s price, and
hence, raise their bids in this auction, which would hurt the bidder if she wins. Thus,
bidders would prefer not to bid in early rounds, and only bid in the last round. In fact,
bidders would delay all bidding until the very last round to avoid pushing other’s bids
upward in later rounds. In other words, last minute bidding occurs in the first auction.
This example is not special, since we can easily generalize the above result to N
bidders and/or to distributions other than uniform, under independent private value
13Note that when b ≥ v1, G(y|v1) is unchanged. Hence, when b ≥ v1, E[V(3)|·] remains unchanged as
well.
8
(IPV). Furthermore, going beyond independent values is also feasible, as shown in the
general model.
II. The eBay Auction
We now present the basic settings of the general model.
A. Basic Settings
We adopt the symmetric affiliated private value setting as in Milgrom and Weber
(1982, 2000) where bidders have unit demand.14 Specifically, we assume that there
are N bidders, M sellers where N > M . There is no discounting and all players are
risk neutral. The entire sequential auction is divided into M stages where one seller
auctions off one item in each stage. The sellers value the items at v0 = 0. Bidders only
want one item and have valuations vi, which are called their types. Types vi are private
and drawn from the symmetric15 joint distribution function F (v1, v2, · · · , vN ) and pdf
f(v1, v2, · · · , vN ), with support V = [v, v]. For cases with a minimum increment s,
we assume v ≥ s > 0 to avoid problems for the lowest type. Types are affiliated.16
The order statistics of {v1, v2, · · · , vN} are {V(1), V(2), · · · , V(N)} where
V(1) ≥ V(2) ≥ · · · ≥ V(N).
To simplify things, we focus on symmetric sequential (perfect Bayesian) equilib-
14Since last minute bidding can be easily explained in a common value setting, we do not embrace the
most general affiliated value model of Milgrom and Weber (1982, 2000).15i.e. for all 1 ≤ i, j ≤ N , F (v1, · · · ,vi, · · · ,vj, · · · , vN ) = F (v1, · · · ,vj, · · · ,vi, · · · , vN ).16See Milgrom and Weber (1982) for the formal definition and properties of affiliation. In particular, for
vectors z and z′, let z ∨ z′ be the component-wise maximum, and z ∧ z′ be the component-wise minimum.
Then, random variables X with joint pdf f(X) is affiliated if, for all z and z′,
f(z ∨ z′)f(z ∧ z′) ≥ f(z)f(z′).
9
rium. Also, we do not consider (weakly) dominated strategies,17 and assume that there
is no waiting cost.18
B. The eBay Auction Rules
The eBay auction has several “non-standard” features that makes it different from
English auctions described in the literature.
First of all, in eBay auctions, we observe out-cry bids instead of drop-out prices.
This is contrary to what is done in the literature. In auction theory, English auctions are
typically modeled as button auctions (with or without reentry), in which an auctioneer
announces prices, and the bidders indicate whether they stay in or not by “pressing on
a button” indicating their participation. Bidders decide which price to drop out and let
go of the button at that price. Drop-outs are observed immediately, and the auction
ends when all but one bidder have dropped out. The remaining bidder wins the item
and pays the current price, which equals to the price when its last rival dropped out. In
the irrevocable button auction, as in Milgrom and Weber (1982), the drop-out decision
is irrevocable so that bidders cannot return once letting go of their button, while in the
button auction with reentry, as in Sergei Izmalkov (2003), bidders are allowed to return
by pressing on the button again. In both settings, dropping out at one’s valuation is a
weakly dominant strategy if bidders are assumed to have private values.19
17Note that asymmetric equilibria arise even in the single unit case if weakly dominated strategies such as
bidding zero is allowed. In particular, having one bidder bidding very high and all others bidding zero is an
equilibrium for a sealed-bid second price auction.18The “no waiting cost” assumption might not hold in internet auctions, but is always assumed in the
auction literature. The key thing needed for our results is that the waiting cost for being available exactly at
the last minute is lower than the gains from “last minute revisions.” Hence, we could approximate this is by
having an automated sniper program bid on your behalf if you are not available.19As discussed in Izmalkov (2003), reentry is introduced to restore efficiency in an interdependent value
setting. All existing efficient equilibrium without reentry can be easily reproduced when entry is allowed.
Hence, in the private setting considered in this section, the equilibrium of the irrevocable button auction can
be reproduced even with reentry.
10
However, neither way described above is suitable to model the eBay auction since
in the eBay auction we only observe bidders’ out-cry bids instead of “drop-outs.”20
Moreover, the interesting phenomenon of “last minute bidding” is specifically the case
where bidders hardly ever call out bids save at the last minute, which is equivalent to
‘almost always staying out” in these button auctions. Therefore, we need to model the
eBay auction in a different way,21 as a type of out-cry auction, in which out-cry bids
instead of drop-outs are observed.
Secondly, eBay auctions utilizes a proxy bidding system where bidders can submit
their reservation prices and let the proxy system bid for them. Bidding on the bidders
behalf, the proxy bidding system will only bid up to the amount which is necessary.
In particular, the system would only overbid the current rival bid (current highest op-
ponent bid) by at most the minimum increment s. When your bid is the standing
bid and another bidder overbids you, the proxy bidding system will revise your bid
to counter it unless the other bidder’s bid exceeds your reservation price. Although
agent-bidding is sometimes observed in physical auction houses as well, the auction
literature has mainly ignored such a possibility. There are two main effects of the
proxy bidding system. On the one hand, it eliminates the possibility of jump-bidding
since the current price only increases up to the rival bid, and hence, eliminates the pos-
sible jump-bid equilibria.22 On the other hand, the proxy bidding system simplifies the
bidding process since it automatically makes counter bids for you immediately when
you are overbid by another bidder. Besides these two main effects, the proxy bid-
20This information issue matters in a repeated setting since we learn bidders’ valuations by observing the
price they drop out. In particular, Milgrom and Weber (2000) showed that an M times repeated irrevocable
button auction where drop-outs are observed would yield exactly the same (ex post) prices in each single-unit
irrevocable button auction since we observe the valuation of the (M + 1)th highest bidder when that bidder
drops out. However, such an equilibrium can not exist if “drop-outs” are not observed.21similar to Ockenfels and Roth (2006), and probably first analyzed by Robert Wilson (1975).22Jump bidding can play a role in signalling bidder valuation, as in Avery (1998). However, such discussion
is beyond the scope of this paper.
11
ding system, when joined with other features, can dramatically change the rules of an
auction as discussed below.
To summarize, the eBay auction has the following bidding procedure (in each
auction stage):
Assumption 1. (Out-Cry Auction with Proxy Bidding)
1. Submitting Proxy Bids: Bidding starts from the reserve price r = v0 at time
t = 0. Bidders can submit a “proxy bid” any time t ∈ [0, T ]. If no bid is
submitted in [0, T ], the auction ends without a sale. Bidder i’s proxy bid at time
t is recorded as bi(t). However, this proxy bid is not revealed immediately. The
(hidden) bidding record of proxy bids at time t is ht ={
bi(t)}
t∈[0,t)
2. High Bidder and its Standing Bid: For any time t ∈ [0, T ], the high bidder,
denoted by bidder k, is the bidder who submitted the highest proxy bid
bk(t1) = maxbi(t)∈ht
bi(t), t1 = arg maxt∈[0,t)
{bi(t)
∣∣∣bi(t) ∈ ht
}.
The second-highest bidder, denote by bidder j, is the bidder who submitted the
second-highest proxy bid
bj(t2) = maxbi(t)∈ht\{bk(t1)}
bi(t), t2 = arg maxt∈[0,t)
{bi(t)
∣∣∣bi(t) ∈ ht \ {bk(t1)}}
.
If there are two or more bidders who submitted their proxy bids, the standing bid
or “current price” St at any time t ∈ [0, T ] is bj(t2), which is the highest proxy
bid that is submitted by a different bidder.23 If there is only one active bidder,
23In practice, eBay actually has a standing bid of
St = min{
bj(t2) + s, bk(t)}
which is the highest proxy bid that is submitted by a different bidder plus an increment unless the difference
between the two proxy bids is smaller than the increment. However, this creates a small possibility that the
eBay auction becomes a first price auction (when the two highest bidders are close), which makes bidding
one’s valuation no longer a dominant strategy even in the single auction case. In fact, with only one auction,
12
the reserve price r. If two bidders submitted the same highest proxy bid, the
bidder who submitted their proxy bid(s) earlier becomes the high bidder, and
the standing bid equals to the tied highest proxy bid. The standing bid St is
common knowledge immediately after time t.24
3. Ascending Proxy Bids: Later proxy bids have to be higher than the bidder’s last
submitted proxy bid25 and higher than the standing bid by a minimum increment
s.26 In other words, if bi(t) is the last proxy bid submitted by bidder i, then its
new proxy bid bi(t) must satisfy
bi(t) > bi(t), and bi(t) > St + s.
Thirdly, the eBay auction has a fixed-time ending rule, which ends the auction at
a pre-set time, commonly known to all bidders, regardless of the bidding behavior
during the course of the auction. This is contrary to the button auction which ends
when all other bidders drop out, and differs from typical out-cry auctions seen in the
the first order condition now requires the bid function to satisfy: (for δ(vi) = vi − β−11 (β1(vi)− s))
β(v) = v − β′(v) · F1(v|v)− F1(v − δ(v)|v)f1(v|v)︸ ︷︷ ︸
shade as if first price
.
24Note that once a proxy bid is overbid, it is automatically revealed through change in the standing bid.
Hence, proxy bids are usually revealed with a delay, except the winning proxy bid, which is never revealed—
not even after the end of the auction.25In other words, a bidder cannot lower its still hidden proxy bid. The only way to lower an existing proxy
bid, is to retract it and re-submit a new one. However, bid retractions are extremely rare in practice, and if
ever happens, is viewed as bad signals. (Note that if bid retraction is costless, the sellers would have a strong
incentive to bid himself, and retract her bid whenever she (accidentally) becomes the high bidder.)26Note that this minimum increment rule does not apply on others’ (hidden) proxy bid, but only on the
standing bid St. Hence, it is possible to submit a proxy bid close to the (hidden) proxy bid. This is quite
different from the continuous time out-cry auction. Another difference is that the ascending rule here only
apply to bids submitted after the standing bid, but not those that were submitted simultaneously.
13
field which end when there are no further bids.27
In other words, we have:
Assumption 2. (Fixed-Time Ending Rule) The auction ends at a fixed time T . The
high bidder wins the item pays the standing bid ST . Note that due to the proxy bidding
system, bidders may submit their reservation price at time t = T . This makes the “last
minute” T as if a sealed bid second price auction.
Last but not the least, since there are many independent buyers and sellers par-
ticipating on eBay, identical items are likely to be sold separately but repeatedly on
eBay. Repeatedness does not drastically change the rules of the auction, except that
the auction house repeats the same single-unit auction M times, each for one item. We
denote each single-unit auction as an “auction stage.” Repetition pushes the eBay auc-
tion away from the single-unit auction framework, and moves towards a multi-unit,
market like structure. Also, it makes bidders not only care about their own private
values, but also what they might need to pay in subsequent auctions.
To see the influence of repetition, we first define the bid history of the repeated
eBay auction:
Assumption 3. (History in the Repeated eBay Auction)
1. Revelation of Last Minute Bids:
We assume that all last minute bids are accepted and ordered from the smallest
to the largest. Hence, no last minute bid would be rejected due to the ascending
rule.28
27As shown below, the ending rule which requires no further bids can be called the “going-going-gone”
rule, and is a key difference between eBay and Amazon auctions.28Note that if the last minute bids do not come in exactly at the same time, information revelation is actually
determined by the the order of these last minute bids. Later bids which are lower then the standing bid (plus
an increment) are rejected by the ascending bid rule, and hence, are not revealed. If we assume the order of
the last minute bids are random, the revelation is also somewhat random. Alternatively, we might think of
endogenously determining the order. If last minute bids are ordered from the smallest to the largest, most
14
2. Public Bidding Record of Past Stages:
At any stage m, the bidding record of the ended auction at stage m′ < m is the
sequence of proxy bids (except the winning one)29
hm′,T ={
bm′i (t)
}t∈[0,T ]
\{
maxt∈[0,T ]
bm′i (t)
}3. Public Bidding Record of Current Stage:
The bidding record of the current stage m, time t ∈ [0, T ] is denoted by the
sequence of standing bids Sm,τ ∈ V ∪ {∅}, as well as the identity of the bidders
iτ ∈ I0 = {0, 1, · · · , nm,t} who submitted proxy bids at τ ∈ [0, t], and can be
transformed into a sequence of proxy bids30
hm,t ={
bmi (t)
}t∈[0,t)
\{
maxt∈[0,t)
bmi (t)
}where bm
i (t) is bidder i’s proxy bid at time t (of stage m).31
4. Bidding History: The entire bidding history up to stage m, time t is defined by
hm,t ={
h1,T , · · · , hm−1,T , hm,t
}last minute bids (except the highest one and those blocked out by the minimum increment requirement) are
revealed. If last minute bids are ordered from the largest to the smallest, then only the second-highest last
minute bid, which determines the final standing bid, is revealed. Other cases are in between. Information
revelation might influence the results. Nevertheless, as shown below, when there are only two items (M = 2),
all information structures result in the same equilibrium.29When the two highest proxy bids are very close, there is a chance that the highest bid is also revealed
(through standing bid Sm′
T ). However, as shown below, this information does not matter in equilibrium when
M = 2.30Amazon actually reveals these proxy bids immediately after they are outbid by other bidders. eBay
initially delayed this to the end of the auction, but then quietly changed to the Amazon way. This is almost
identical, since we can almost always derive the outbid proxy bids from the change in the standing bids.
However, there might be a subtle difference if the proxy bids are too close. In this paper, we abstract away
from these subtleties.31Note that the winning proxy bid is not revealed. Also, bids submitted at the same time t might be rejected
due to the ordering of these simultaneous bids and the ascending rule. To avoid this, we can simply assume
simultaneous bids are ordered from small to large so no bid is rejected.
15
Hence, for any stage m, the past history space is
Hm =m−1∏m=1
[ ∏t∈[0,T ]
((V ∪ {∅})⊗ I0
)n]
t
m
,
and the history space of the current stage is
Hc =∏
t∈[0,T ]
[((V ∪ {∅})⊗ I0
)n]t,
so that the entire history space is H = HM−1 ×Hc.
Note that the symmetric bidding strategy for a bidder with valuation v is a function
β : V ×H =⇒ {∅} ∪ V.
In particular, the bidding strategy bidder i with valuation vi at time t of stage m (after
history hm,t) is
βi(vi, hm,t) =
∅ (not bid)
bi ∈ V (bid bi)
Therefore, the strategy space is V ×H × (V ∪ {∅}), which is potentially huge.
To summarize, an eBay auction is defined to be an out-cry auction with proxy
bidding and a fixed-time ending rule. If there are more than one item sold, the same
eBay auction is conducted repeatedly using the same rules. The combination of these
non-standard features can produce unexpected results. For example, the proxy bidding
system at the fixed ending-time works like a one-shot sealed bid second price auction.
This makes the eBay auction more closer to a sealed bid second price format, and
combining with the last minute bidding phenomenon, motivated Bajari and Hortacsu
(2003) to analyze eBay coin auctions as if they were actually sealed bid second price
auctions.
We first focus on the eBay auction and address the overall effect of all of these
non-standard features, and discuss other possible formats and the effect of individual
features in the discussion section.
16
C. The “Puzzle” of Last Minute Bidding
The last stage of an eBay auction is simply a single-unit auction. In this environment,
we can easily find the unique equilibrium by (weak) dominance. To do this, first
observe that any bidding strategy that does not “bid up to one’s valuation,” i.e. the
resulting bidding sequence does not end with a final bid between vi − s and vi, is a
dominated strategy. This is because one can modify that strategy by adding another
bid of one’s valuation to the end of that sequence. Moreover, given that everybody else
will “bid up to their valuation,” a bidder will find bidding her valuation at time zero
is (weakly) dominant since she might be “blocked out” from bidding by the minimum
increment rule. Therefore, we have that in the last stage of the eBay auction, the unique
equilibrium is characterized by bidders submitting proxy bids bi(0) = vi at t = 0, and
cease to bid after the initial bids are submitted. Hence, the unique equilibrium of
eBay auction’s final stage is the same as that in the (irrevocable) button auction and is
efficient.
Note that in this equilibrium, all bidders submit a proxy bid equal to their valua-
tions at the time 0 and the ending rule has no bite.
Also note that when there is only one auction (M = 1), all bidders submit proxy
bids equal to their valuation at time 0, not at the last minute (time T ) observed in
practice. This is why “sniping,” or last minute bidding, is not explained by (and even
contradicts) standard (single-unit) auction literature. Specifically, we have the follow-
ing “puzzle” of last minute bidding:
Proposition 1. (eBay Auction)
Assume that there is only one item sold under the eBay auction rule. The unique
equilibrium is characterized by bidders submitting proxy bids bi(0) = vi at t = 0, and
cease to bid after the initial bids are submitted. Hence, the unique equilibrium of the
(single-unit) eBay auction is the same as that in the (irrevocable) button auction and
is efficient.
17
Proof. See Appendix.
Hence, in order to explain last minute bidding, Ockenfels and Roth (2006) add an
exogenous probability p that the last minute bid submitted at time T is not successfully
transmitted due to internet congestion or disconnection. They construct a “collusive”
equilibrium in which bidder only bid at the last minute hoping for the small probabil-
ity of gaining a lot when others’ bids are lost. However, it is difficult to imagine how
numerous online bidders who never knew each other could jointly collude. What is
more, such equilibrium relies heavily on appropriate parameters such as the distrib-
ution of values, the probability p and the number of bidders. In fact, Ockenfels and
Roth (2006) only gave an example where there are only two bidders with identical
values as proof for the possibility of such collusive equilibrium, instead of specifying
exact conditions for it to happen.32
We take an alternative approach, and consider the situation where many identical
items are sold in the same (eBay) auction house, but each auction ending at a different
time. This adds repetition to the model.
III. Equilibrium of the Repeated Ebay Auction
Note that the eBay auction has the sealed-bid second price auction as a special cases
when T = 0. In other words, the (repeated) eBay auction with T = 0 is equivalent to
a (repeated) sealed-bid second price auction. We first consider this special cases as a
benchmark.32Ariely et al. (2005) has another numerical example, but still not a general result. Moreover, Ariely et
al. (2005) show that subjects still bid at the last minute even in sessions where the probability p was zero,
contrary to their own “collusive” theory. Hence, they explore other non-equilibrium explanations such as best
responding to incremental bidders.
18
A. The Repeated Sealed-bid Second Price Auction
When T = 0, which eliminates all but the last round of bidding, the eBay auction
reduces to a sequence of sealed-bid second price auctions. The equilibrium of this
auction can be found through backward induction. In the last stage, as shown above,
bidding one’s valuation (at T = 0) is the (weakly) dominant strategy. Moving back to
the early stage, in this second price setting, bidders would bid their expected prices of
the next item, conditional on themselves being the tied winner for this one, since this
is the “willingness-to- pay” for this auction stage. Hence, we have:
Theorem 1. (APV, Any Information Structure, 2 Stages)
Assume that two items are sold under the repeated sealed-bid second price auction, i.e.
the eBay auction rule with T = 0, and s = 0. Then, no matter how many last minute
bids are revealed after the auctions (arbitrary on information revelation assump-
tion 1.), a symmetric equilibrium is determined by bidding β1(vi) = E[V(3)
∣∣∣V(1) =
vi, V(2) = vi
]in stage 1, and β2(vi) = vi in stage 2. Moreover, expected revenue
is weakly higher in the last stage, which equals to E[V(3)
]. In particular, when val-
uations are independent, we have revenue equivalence such that stage 1 and stage
2 yield the same expected revenue, and total revenue equals to that of a multi-unit
uniform price (Vickrey) auction selling all items simultaneously.
Proof. See Appendix.
Now we consider the general case where T is positive and M = 2, in which two
identical items are sold using the repeated eBay auction rule. First, we show there
exists a last minute bidding equilibrium where everyone jumps in and bids at the last
minute. Then, we can consider uniqueness of such equilibrium.
B. Existence of Last Minute Bidding Equilibrium
To get existence, we need several lemmas about affiliation and information updating.
First, note that a bidding strategy without a last minute bidding feature is weakly
19
dominated by the bidding strategy which is almost identical except “adding” a last
minute bidding clause, and if a bidder already bid her valuation in some early stage,
she would not bid at the last minute even with such a clause. Another thing to note is
that revealing information cannot revise the expected prices lower in these eBay open
auctions.33 To get this result, we need properties about affiliation, which is mainly
covered in Milgrom and Weber (1982), except the self affiliation property stated below.
To summarize, we have the following three properties of affiliation:
Lemma 1. Suppose v1, v2, · · · , vN are affiliated, and we are in the repeated eBay
auction game, then
1. (Self-affiliation) For Y = v1, Y, v1, v2, · · · , vN are affiliated.
2. (Last minute bidding clause) For M = 2 and s = 0, a bidding strategy β(vi, hm,t)
such that β1(vi, h1,T ) = ∅ is weakly dominated by
β(vi, hm,t) = β(vi, hm,t), for all vi and (m, t) 6= (1, T ).
= E[V(3)
∣∣V(1) = vi, V(2) = vi, h1,T
], if (m, t) = (1, T ).
3. (Belief updating) After any history hm,t of the sequential eBay auction where
V1, · · · , Vn are still affiliated, for a > v,
E[min{V(M+1) + s, Vi}
∣∣∣V(m) = vi, V(m+1) = vi, Vi = vi, a < Vj < v]
≥ E[min{V(M+1) + s, Vi}
∣∣∣V(m) = vi, V(m+1) = vi, Vi = vi, v < Vj < v].
Proof. See Appendix.
Now we show existence by constructing a last minute bidding equilibrium where
all bidders wait and jump in at the last minute, t = T , bidding the repeated sealed-bid
33Note that Milgrom and Weber (1982) provided a similar result for the seller in the general affiliated value
(AV) settings, stated as Theorem 8: In the second price auctions, publicly revealing the seller’s information
raises revenue (weakly).
20
second price auction equilibrium characterized in the previous section, according the
players’ ex ante information in the beginning of this stage.34
Theorem 2. (Existence of the Last Minute Bidding Equilibrium)
Assume that there are two items sold under the repeated eBay auction rule, and s ≥ 0.
Then, there exists a (symmetric) last minute bidding equilibrium where
β1 =
∅ t < T
E[V(3)
∣∣V(1) = vi, V(2) = vi
]t = T
β2 =
vi t = 0
∅ t > 0.
and the out of equilibrium beliefs are
1. If at some time t, St moves to become S0 = a > v, and at the same time bidder
j1 replaces bidder j2 as the high bidder, then infer that v ≥ vj1 , vj2 ≥ S0 > v.
2. If at some time t, St = S0 ≥ 0, but S0 ≤ v, then infer that v ≥ vj ≥ a > v for
some arbitrary given a > v, for all vj that have submitted a bid.
Proof. See Appendix.
Note that if s > 0, this equilibrium might not sustain if the minimum increment s
is charged on simultaneous bids since this would create a “first price” situation when
valuations are close.35
34Since there is no information update, the interim information h1,T , h2,T is the same as the ex ante
information h1,0, h2,0.35When simultaneous bids are charged with the minimum increment, and hence, the winning price being
min{v(3) + s, vi}, bidders have to consider the possibility of falling into the “first price auction” range where
v(1) − v(2) < s. Hence, the first order condition becomes
β1(vi) = E[min{Y2 + s, Vi}
∣∣Y1 = vi, Vi = vi
]− β′
1(vi) ·F1(vi|vi)− F1(vi − δ(vi)|vi)
f1(vi|vi)
where δ(vi) = vi − β−11 (β1(vi)− s). In other words, when there is a chance of determining the winning
price (when the auction is close), bidders shade their bids from their conditional expectations to account for
this “first price” possibility.
21
This is because the ascending rule applies only to “earlier” or “later” proxy bids,
so these simultaneous bids would not be blocked out. Moreover, one deviation would
not crowd others out since the standing bid is only updated to be the reserve price
r = v0 = 0, and the lowest type is v ≥ s > 0.
C. Uniqueness of the Last Minute Bidding Equilibrium
Now we attempt to argue for the uniqueness of this last minute bidding equilibrium.
To do this, we eliminate bidding strategies that are weakly dominated. Recall that
focusing on symmetric equilibrium already rules out asymmetric equilibria where one
bidder bids very high and others bid zero, which rely on weakly dominated strategies.
We can now state our uniqueness theorem:
Theorem 3. (Uniqueness of the Last Minute Bidding Equilibrium)
The last minute bidding equilibrium is the unique symmetric equilibrium if we consider
only undominated and monotonic strategies for the repeated eBay auction rule. In
other words, people simply play the last minute bidding equilibrium strategy where
everybody jumps in at the last minute except in the last stage, and hence, bid according
to their ex ante information, obtaining the repeated sealed-bid second price auction
outcome.
Proof. See Appendix.
Intuitively, since bidding early reveals information and increases your opponents’
conditional expectation, and hence, increases their last minute bids, you would rather
withhold bidding until the last minute.
According to the theorem, last minute bidding can be rationalized even in the pri-
vate value case, and we have shown that with repeated auctions, bidders’ reservation
prices in the first period are actually correlated across bidders since the conditional
expectations of v(M+1), which are their ex ante reservation prices in the early stages,
depends on not only one’s type but others’ types as well.
22
Note that there are two interesting empirical implications of this theorem. First of
all, if the auctions observed are repeated, private values(PV) and common values(CV)
would be difficult to distinguish empirically, since even if the true valuations were
private values, the bids would still depend on other bidders’ information in early auc-
tions. Unfortunately, this is typically the case since similar items are auctioned off
over and over across time, and in fact, we need to observe similar auctions performed
repeatedly to identify our empirical estimates.
Moreover, though in principle we should find no winner’s curse in private value
auctions, there might still be winner’s curse in repeated private value auctions. To be
precise, in the theoretical analysis above, bidders avoid the winner’s curse since all of
the expectations were conditional on Vi = V(m+1) > V(M+1). However, if people do
not condition on that, they would still suffer the winner’s curse.
D. Revenue Ranking and Efficiency
Since the symmetric equilibrium we consider assumes monotonic strategies, the win-
ners of stage 1, · · · ,M have valuations v(1), · · · , v(M), respectively. Therefore, the
outcome is efficient. Moreover, sellers revenue differ between stages. In fact, from the
proof of Theorem 1, we have shown that
Theorem 4. (Revenue Ranking; APV, M = 2)
Suppose there are two items sold (M = 2) under the eBay auction (or sealed-bid
second price auction) rule, and s = 0. Then, the expected revenue of each stage is
E[β1(V(2))
]and E
[V(3)
]where E
[β1(V(2))
]≤ E
[V(3)
], and hence, sellers’ total
expected revenue is lesser or equal to 2 · E[V(3)
], which is what they would earn if
they combined the items and held a joint sealed-bid auction where the winning price
is the highest loser’s bid.
Thus, with affiliation, last minute bidding does not hurt efficiency, but might leave
room for sellers to improve in terms of revenue. In fact, if the first seller could join
with the second seller to host a multi-unit second auction in which the first and second
23
highest bidder win an item and pay the third highest price, the first seller’s expected
revenue would increase, while the second seller would remain indifferent.
Although revenue equivalence generally breaks down in the APV case,36 with in-
dependence, we still have revenue equivalence, as shown in the proof of Theorem 1. In
other words, under independence, the repeated eBay auction yields the same expected
revenue for each stage, and sellers are indifferent between selling in different stages.
Therefore, under independent private value, last minute bidding is not “bad” in the
sense of both efficiency and revenue.37 We restate it as
Corollary 5. (Revenue Equivalence; IPV, M = 2)
Suppose vi are independent, there are two items sold (M = 2) under the eBay auction
(or sealed-bid second price auction) rule, and s = 0. Then, the expected revenue of
each stage is E[V(3)
], and the sellers’ total expected revenue is 2 · E
[V(3)
].
This result is a special case of Weber (1983) proved the general revenue equiva-
lence theorem for the IPV case.38
36Milgrom and Weber (1982) rank different auction rules according to their expected revenue in the general
setting. They argue that irrevocable button auctions are widely used due to its higher expected revenue,
compared to the sealed-bid first price or second price auctions.37The revenue ranking for more than two items under affiliated private value is partially discussed in
Joseph Tao-yi Wang (2006). In particular, the sealed-bid second price auction (when the highest two bids are
revealed) yields lower revenue for earlier auctions than later ones. A similar last minute bidding equilibrium
would exist in the eBay auction as well, but uniqueness may not hold.38Specifically, Weber (1983) has:
Proposition. (Revenue Equivalence; IPV; Weber)
Suppose vi are independent. For any auction rule that assumes that in equilibrium, the M highest types
win the auction for sure, and the lowest type v expects to get 0, the total seller(s) expected revenue is M ·
E[V(M+1)
].
24
IV. Discussion: Other Auction Formats
After presenting the main results of the eBay auction, we now turn to discuss other
possible auction formats and demonstrate why last minute bidding behavior cannot be
easily explained. As discussed earlier, the eBay auction is an out-cry auction with two
special features: the fixed-time ending rule and the proxy-bidding system. We may
consider other types of out-cry auctions that have only one (or neither) of the these
special features.
Without any special features, we have the simple out-cry auction. This is typically
the way English auctions are carried out in the field, though the auction literature
usually models it as a button auction (with or without re-entry).39 Adding a fixed
ending time to the simple out-cry auction, we have a fixed-ending out-cry auction.
On the other hand, if we do not have a fixed ending rule, but add a proxy-bidding
auction, we obtain an Amazon auction, named after the online auction house which
uses this format. We may summarize the four different variants of the out-cry auction
as follows:
Proxy-bidding System
Yes No
Fixed-time Yes eBay auction Fixed-ending out-cry auction
Ending Rule No Amazon auction Simple out-cry auction
We now proceed to discuss these formats.
A. The Simple Out-cry Auction
As discussed earlier, there are two ways to model the English auction in the auction
literature. One is the irrevocable button auction analyzed by Milgrom and Weber
(1982), and the other is the button auction with reentry analyzed by Izmalkov (2003).
39Indeed, we show that there exist a button auction equilibrium in the simple out-cry auction which resem-
bles the equilibrium of a button auction.
25
In both cases, the auctioneer announces prices, and the bidders indicate whether they
stay in or not by pressing on a button indicating their participation.
However, as a by product of discussing eBay auctions, we model the English auc-
tion in a third way, as an out-cry auction with the following bidding rules:
Assumption 4. (Out-cry Bidding Procedure)
1. Submitting Bids: Bidding starts from the reserve price r = v0 at time t = 0.
Bidders can bid at any time t ∈ [0, T ]. If no bid is submitted in [0, T ], the
auction ends without a sale. Bidder i’s bid at time t is recorded as bi(t). This
bidding record is available immediately after time t. Hence, the bidding record
at time t is ht ={
bi(t)}
t∈[0,t)∈ H where H =
∏t∈T (V ∪ {∅})n is the history
space.
2. Bidding Sequence: Bids are ordered according to its submission time t. If two
or more bids enter at the same instance t, they are randomly ordered. i.e. the
bidding sequence is b1 = bi1(t1), b2 = bi2(t2), · · · , bn = bin(tn), · · · , where
tk ≤ tk+1.
3. Ascending Bids: Later bids have to be higher than their predecessors by a mini-
mum increment s. In other words, if the bidding sequence is b1, b2, · · · , bn, · · ·
, then bk+1 ≥ bk + s for all k ∈ N.40 We denote the last and highest bid (at time
t) as the standing bid or current price St.
4. Symmetric Bidding Strategies: The bidding strategy for a bidder with valuation
v is a function
β : V ×H =⇒ {∅} ∪ V
40If two bids are submitted at the same instance t, the (randomly selected) latter bid bik+1(tk+1) would be
rejected and not recorded if it does not satisfy this minimum increment requirement. However, this is fine
since the bidder can always resubmit another bid at t′ > t if she has valuation vik+1 ≥ St + s. Only if there
is a fixed last minute to submit bids or if the minimum increment is larger than the difference between vik+1
and St would this be an issue. Nevertheless, if |vik+1 −St| < s, the potential efficiency loss is at most s, and
vanishes as s → 0.
26
The bidding strategy bidder i with valuation vi at time t (after history ht) is
βi(vi, ht) =
∅ (not bid)
bi ∈ V (bid bi)
Just as in reality, there is an auctioneer soliciting bids several times before an-
nouncing that the item is sold, we have the going-going-gone ending rule:
Assumption 5. (going-going-gone ending rule) The auction ends if no further bids
came in during the “going-going-gone” period(tn, tn+T
]after the last bid was sub-
mitted at tn. The high bidder wins the item and pays the standing bid Stn . If tn +T >
T , then the auction is automatically extended to time T ′ =(tn + T
).41 The bidding
history of this ended auction is the finite42 bidding sequence h ={
bik(tk)}n
k=1where
bk(·) is bidder ik’s bid at time tk.
A simple out-cry auction is an out-cry auction with a going-going-gone ending
rule.43 When there is only one auction conducted, i.e. M = 1, we may characterize
the equilibrium of the simple out-cry auction by the following lemma:
Lemma 2. Assume that there is only one item sold (M = 1) under the simple out-
cry auction rule. If it is common knowledge that s < v(1) − v(2), an equilibrium is
characterized by a finite bidding sequence {b1, · · · , bn} with the bidder with valuation
v(1) bidding the last bid bn ∈((v(2) − s), (v(2) + s)
], and all other bidders cease to
bid during the going-going-gone period(tn, tn +T
). Hence, the outcome is efficient,
and as s → 0, the winning price converge to the winning price in the (irrevocable)
button auction.
Proof. See Appendix.
41Actually, automatic extension is not necessary if T is large enough, just as in practice the auctioneer
usually has enough time to solicit for new bids. What is crucial is the “going-going-gone” ending rule.42Since we have a minimum increment s, the sequence of bids is finite as long as the winning bid is not
infinity.43This is in contrast to the fixed-ending out-cry auction discussed in the next section.
27
In fact, this result may be the very reason previous literature modeled English
auctions directly as button auctions, even though they are actually conducted in the
field as simple out-cry auctions.
In any case, if the valuations have a smooth pdf, we can actually specify the equi-
librium as follows:
Proposition 2. (Simple Out-cry Auction)
Assume that there is only one item sold (M = 1) under the simple out-cry auction
rule. If f2, the pdf of V(2)(= Y1), satisfies
f2(y + η)f2(y)
≤ 4 for all y ∈ [v, v] and η ∈ [s, 3s],
there exists an button auction equilibrium (BAE) characterized by a series of bids
bi(t) = s, 2s, · · · , ks, · · · , at t = T , 2T , · · · , kT , · · · ,
as long as vi ≥ ks, for T ≤ T .
As s → 0, the above assumption becomes less and less binding, and the outcome
approximates an irrevocable button auction.
Proof. See Appendix.
Intuitively, the button auction equilibrium produces an outcome that mimics the
price clock in a button auction. Hence, there is little to gain by deviating from the
equilibrium strategy. However, smoothness is required here so bidders cannot gain by
jump bidding and blocking opponents with valuation right above her. This requirement
is less and less binding as s → 0, and is trivially satisfied when s = 0.
Note that when M ≥ 2, the equivalence result between the (irrevocable) button
auction and the continuous-time out-cry auction may not hold.44
44This is because the simple out-cry auction does not require all other bidders to counter bid, which is
equivalent to not dropping out in the (irrevocable) button auction, but just at least one counter bid. Hence, two
bidders might engage in a bidding war against each other which is unnecessary since one of them could stop
28
B. The Fixed-ending Out-cry Auction
Without the proxy-bidding system, the fixed-time ending rule would make the “last
minute” work like a first price auction. Hence, this would make the out-cry auction
end with one last round of sealed bid first price auction. Such an auction format
would be similar to the Anglo-Dutch auction proposed by Kenneth G. Binmore and
Paul Klemperer (2002) for the British 3G telecom auction, except that the last round
of sealed bid first price auction is triggered by a fixed ending time. We refer this
auction format as the Fixed-ending Out-cry Auction, which is literally the eBay auction
without proxy-bidding.
If there is only one item sold (M = 1), the fixed-ending out-cry auction has a
button-auction equilibrium (after re-mapping bidding time into the fixed time period),
just as the simple out-cry auction. In other words, we have the following proposition:
Proposition 3. (Fixed-ending Out-cry Auction)
Assume that there is only one item sold (M = 1) under the fixed-ending out-cry auc-
tion rule. If f2, the pdf of V(2)(= Y1), satisfies
f2(y + η)f2(y)
≤ 4 for all y ∈ [v, v] and η ∈ [s, 3s],
there exists a button auction equilibrium (BAE) characterized by
bi(t) = s, 2s, · · · , ks, · · · ,Ks, at t = T , 2T , · · · , kT , · · · ,KT
where K =[
vs
]+ 1, the closest (but higher) integer to v
s , and T ≤ TK , guaranteeing
bidding would be “complete” before the fixed ending time. as long as vi ≥ ks, for
T ≤ T .
counter bidding, wait, and bid on subsequent auctions. However, this creates incentives for high valuation
bidders to jump bid, eliminating the chance for multiple opponents to counter bids revealing their relatively
high valuation, and hence, deter other bidders from counter bidding. Such jump-bid equilibria rationalizes
the imposed game structures in the jump-bid literature such as Avery (1998). However, further discussion is
beyond the scope of this paper.
29
As s → 0, the above assumption becomes less and less binding, and the outcome
approximates an irrevocable button auction.
The proof is identical to that in the previous section.
One might speculate that a last minute bidding equilibrium could exist in this auc-
tion format where bidders bid as if it were a sealed-bid first price auction. However, to
demonstrate the difficulty of explaining last minute bidding with the fixed-time ending
alone, it is suffice to show that a button auction equilibrium exists in the fixed-ending
out-cry auction. Further consideration is beyond the scope of this paper.
C. The Amazon Auction
As shown earlier, the eBay auction has several non-standard features that makes it
different from the single out-cry auction considered in the previous section. First, it
has a proxy bidding system where you can submit your reservation price and let the
proxy system bid for you. Moreover, the eBay auction has a fixed-time ending rule,
and thus, the proxy bidding system at the fixed ending-time works more like a sealed
bid second price auction.
On the other hand, as reported in Roth and Ockenfels (2002), Amazon.com has
its own auction site which conducts auctions similar to eBay, which also has a proxy-
bidding system but preserving the “going-going-gone” ending rule of the simple out-
cry auction. We refer this auction format, which consists of proxy bidding and the
going-going-gone ending rule, as the Amazon auction.
For the case where only one item is sold, or M = 1, we can characterize the unique
equilibrium of the Amazon auction with the following proposition:
Proposition 4. (Amazon Auction)
Assume that there is one item sold (M = 1) under the Amazon auction rule and
s > 0. Then, the unique equilibrium is characterized by bidders submitting proxy bids
bi(0) = vi at the beginning of the auction, and cease to bid after the initial bids are
30
submitted. Hence, the unique equilibrium of the Amazon auction is the same as that in
the (irrevocable) button auction and is efficient.
Note that in this equilibrium, all bidders submit a proxy bid equal to their valua-
tions at the time 0 and the automatic extension rule has no bite. Hence, the proof is
identical to that of Proposition 1.
In conclusion, the equilibrium for each auction format is given by Proposition 1
through 4, and we may summarize equilibria under the four different variants of the
out-cry auction as follows:
Proxy-bidding System
Yes No
Fixed-time Yes (eBay) Bid at t = 0 (FEOC) Button Auction Eq.
Ending Rule No (Amazon) Bid at t = 0 (SOC) Button Auction Eq.
Moreover, since bidding at time 0 is dominant when there is proxy bidding, last
minute bidding is not generated by the fixed-time ending rule per se. Nevertheless,
without the fixed-time ending rule, there is no “last minute” to bid on! Therefore,
neither proxy bidding, nor a fixed ending time, can induce last minute bidding by
itself. In fact, it is only with both of them, and the fact that there are multiple units
sold which generates last minute bidding behavior.
V. Conclusion
In this paper, we rationalize last minute bidding by adding (at least) another identical
auction. In the repeated eBay auction model, we prove that the last minute bidding
equilibrium, in which bidders only bid at the last minute, is the unique symmetric
sequential equilibrium in weakly undominated and monotonic strategies. Moreover,
bidders do not always bid their valuations. In fact, a bidder’s maximum willingness
to pay for this stage depends on on others’ types. This introduces a “common value”
component to the private value environment and leads to bid shading in the first auc-
tion. Nonetheless, even though high valuation bidders efficiently win the items, under
31
strict affiliation, expected revenue is lower in the first auction than in the last, and
the seller of the first item can increase her revenue by running a multi-unit auction
together with the second seller. Such revenue gains can be viewed as potential syner-
gies for multi-unit auctions, and the foregone revenue to “auction first” highlights the
opportunity cost of coordination failure.
Mathematical Appendix
Proof of Proposition 1. We first show that not bidding up to one’s valuation is (weakly)
dominated. Then, after eliminating these dominated strategies, it suffice to show that
for bidder i, submitting a proxy bid of bi(t) = vi at t = 0 is a (weakly) dominant
strategy.
First, suppose bidder i’s strategy leads to the submission of several proxy bids,
Bi(ht1) < Bi(ht2) < · · · < Bi(htn), and cease to bid after tn ≤ T . Then the final
proxy bid Bn must be in (vi − s, vi]. Otherwise, adding a proxy bid Bi(htn+1) = vi
at time tn+1 ≥ tn would dominate the proposed strategy. Since tn ≤ T , there exists
tn+1 between tn and T , the ending period of the auction.45 Hence, not bidding up to
one’s valuation is dominated.46
Due to the proxy bidding system, submitting a proxy bid bi(h0) = vi at time
t = 0 would yield the same outcome if there does not exist two other bidders who
have valuations vj , vk ∈ (vi − s, vi), or if Bi(htn) = vi. However, by submitting a
proxy bid equal to one’s reservation price one strictly gains if there exists bidders j, k
who have valuations vj , vk ∈ (vi − s, vi), and hence, submit proxy bids bj = vj , and
bk = vk before bidder i. This is because the standing bid St will become either bj
or bk, and is in (vi − s, vi), effectively “blocking out” bidder i due to the ascending
auction rule.45If tn = T , we simply change the last bid so that Bi(htn
) = vi.46This eliminates possible asymmetric equilibria such as having all but one bidder bidding zero according
to some early bidding signal.
32
Since we assume continuous types with full support, there is a positive probability
of having such opponents. Therefore, submitting a proxy bid of bi = vi at time t = 0
weakly dominates the proposed strategy.
Thus, due to the ascending bidding rule, bidders race to bid early, and end up all
bidding at time t = 0.47
Proof of Theorem 1. In the last auction, or stage 2, we have the standard case in which
the remaining bidders bid their valuation β2 = vi and the item is sold at price v(3) since
there are (N − 1) bidders left.
This can be shown by solving the maximization problem of bidder (of type) i
choosing her bid bi, given other bidders’ (identical) strategies b∗−i. Let b(2) be the
second-highest bid, and let the history be hm = hm,0, then the maximization problem
is
maxbi
u(bi, b−i)
= Pr(bi > b(2)) ·(vi − E
[b(2)
∣∣∣bi > b(2), h2
])Suppose the (monotonic) equilibrium bid function is β∗
2(v), we may transform the
maximization problem from choosing bid bi into choosing to “pretend to be type x” in
which β∗2(x) = bi. Given other bidders’ strategy β∗
2(v−i), the maximization problem
becomes:
maxx
u(x, v−i) = Pr(x > v(2)) ·(vi − E
[β∗
2(V(2))∣∣∣x > V(2), h2
])= F−i(x|vi, h2) · vi −
∫ x
vβ∗
2(y1)f−i(y1|vi, h2)dy1
where y1 is the highest type among the bidders other than bidder i, F−i(y1|vi, h2) is
its conditional distribution function given vi, h2, and f−i(y1|vi) is its conditional pdf
given vi, h2. Then, the first order condition is (at x = vi)
∂u
∂x= f−i(x|vi, h2) · vi − β∗
2(x)f−i(x|vi, h2) = 0
47In practice, this corresponds to bidders submitting their bid when they first see the auction.
33
In fact,
∂u
∂x=
(vi − β∗
2(x))· f−i(x|vi, h2)
≥ 0 if x < vi
≤ 0 if x > vi
Therefore, β∗2(vi) = vi maximizes bidder i’s expected payoff.
Knowing this, in the next-to-last stage, or the first stage, the N bidders can cal-
culate the expected winning price of the last stage, and that would be the maximum
amount bidder i is willing to pay at this stage. If the expected winning price in this
stage is more than that of the next stage, they would rather wait and bid in the next
stage. This can be shown by consider the utility of bidder i with valuation vi, but
bidding as if she were x in auction 1, and knowing she would bid as if she were vi in
auction 2.48 Let Y1 > Y2 > · · · > Y(N−1) be the order statistics of the valuations of
the (N − 1) bidders j 6= i, and
Y1
∣∣∣∣V
∼ f1(y1|vi), Y1, Y2
∣∣∣∣V
∼ f2(y1, y2|vi)
then
u(x|vi) =∫ x
v
[vi − β1(y1)
]f1(y1|vi)dy1
+∫ v
x
∫ min{vi,y1}
v
[vi − y2
]f12(y1, y2|vi)dy2dy1
Hence,
∂u
∂x=
[vi − β1(x)
]f1(x|vi)−
∫ min{vi,x}
v
[vi − y2
]f12(x, y2|vi)dy2 = 0
at x = vi if
β1(vi) =1
f1(vi|vi)
∫ vi
vy2f12(vi, y2|vi)dy2
= E[Y2
∣∣Y1 = vi, Vi = vi
]= E
[V(3)
∣∣V(1) = vi, V(2) = vi
]48This is by subgame perfection.
34
Moreover, for x < vi, min{vi, x} = x, and
∂u
∂x= vi
[f1(x|vi)−
∫ x
vf12(x, y2|vi)dy2︸ ︷︷ ︸
f1(x|vi)
]− β1(x)f1(x|vi) +
∫ x
vy2f12(x, y2|vi)dy2
= −β1(x)f1(x|vi) +∫ x
vy2f12(x, y2|vi)dy2 ≥ 0
since for x < vi,
β1(x) = E[Y2
∣∣Y1 = x, Vi = x]
≤ E[Y2
∣∣Y1 = x, Vi = vi
]=
1f1(x|vi)
∫ x
vy2f12(x, y2|vi)dy2
On the other hand, for x > vi, min{vi, x} = vi, and
∂u
∂x= vi
[f1(x|vi)−
∫ vi
vf12(x, y2|vi)dy2︸ ︷︷ ︸
R xvi
y2f12(x,y2|vi)dy2
]− β1(x)f1(x|vi) +
∫ x
vy2f12(x, y2|vi)dy2
= −β1(x)f1(x|vi) +∫ x
vy2f12(x, y2|vi)dy2 ≥ 0
since for x > vi,
β1(x) = E[Y2
∣∣Y1 = x, Vi = x]
≥ E[Y2
∣∣Y1 = x, Vi = vi
]=
1f1(x|vi)
∫ x
vy2f12(x, y2|vi)dy2
Thus, the unique symmetric perfect Bayesian Nash equilibrium for this stage is to
bid
β1(vi) = E[Y2
∣∣∣Y1 = vi, Vi = vi
]= E
[V(3)
∣∣∣V(1) = vi, V(2) = vi
].
The item is won by the bidder with the highest valuation among the N bidders, which
has valuation v(1) and pays the second-highest bid (submitted by the bidder with val-
uation v(2)):
β1(v(2)) = E[V(3)
∣∣∣V(1) = v(2), V(2) = v(2)
].
35
Note that the expected revenue of stage 1 is, by affiliation,
E{
β1(V(2))}
= Ev(1),v(2)
{E
[V(3)
∣∣∣V(1) = v(2), V(2) = v(2)
]}≤ Ev(1),v(2)
{E
[V(3)
∣∣∣V(1) = v(1), V(2) = v(2)
]}= E
[V(3)
].
Hence, for M = 2, the revenue of the first auction is lesser or equal to that of the
last auction. In particular, if valuations are independent, we have revenue equivalence,
since
E[V(3)
∣∣∣V(1) = v(2), V(2) = v(2)
]= E
[V(3)
∣∣∣V(1) = v(1), V(2) = v(2)
].
In fact, when M = 2, the information structure plays no role since the last auction is
the standard case where the dominant strategy for everybody is to bid their valuation.
Proof of Lemma 1. We prove the three properties one-by-one:
1. For v = (v1, · · · , vn), v′ = (v′1, · · · , v′n), and joint pdf f(v1, · · · , vn), we have
(by affiliation)
f(v ∨ v′)f(v ∧ v′) ≥ f(v)f(v′).
Consider z = (y, v) = (y, v1, · · · , vn), v′ = (y′, v′) = (y′, v′1, · · · , v′n), and
g(z) = g(y, v1, · · · , vn) = f(v1, · · · , vn) if y = x1
= 0 if y 6= x1
Then we need to show that g(z ∨ z′)g(z ∧ z′) ≥ g(z)g(z′).
For x1 = y and x′1 = y′,
g(z ∨ z′)g(z ∧ z′) = f(v ∨ v′)f(v ∧ v′)
≥ f(v)f(v′) = g(z)g(z′)
For x1 6= y or x′1 6= y′, g(z)g(z′) = 0 and the inequality is trivial.
36
2. Since there is no waiting cost, coming back at t = T to bid (according to the
then current history hm,T ) is costless. Also, since M = 2, we need not worry
about information updating in the second and last stage. Hence, there is no loss
of submitting a bid.
However, if there were information revealed before the last minute T , bidders
can gain by bidding at T , acting accordingly.
3. Since v1, · · · , vn are affiliated, for s = vj , v1, · · · , vn, s are affiliated by part 1.
Moreover, consider the order statistics of v1, · · · , vn, namely v(1), · · · , v(n).
Then v(1), · · · , v(n), s are affiliated as well since their joint pdf (conditional on
s = vj) is
n! · g(v(1), · · · , v(n), s|s = vj) · 1v(1)≥···≥v(n).
Hence, for min{V(M+1) + s, Vi}, a nondecreasing function of V1, · · · , Vn, we
may apply Theorem 5 of Milgrom and Weber (1982)49 to obtain that
E[min{V(M+1) + s, Vi}
∣∣∣V(m) = vi, V(m+1) = vi, Vi = vi, Vj ≥ a]
is nondecreasing in a.
Therefore we have established that any bid submitted would update the con-
ditional expectation upwards (or unchanged), if the inferred belief is that only
bidder types higher than a > v would do such action.
49Milgrom and Weber (1982) stated several important properties of affiliation. The one used here is
Theorem 5. Let Z1, ·, Zk be affiliated and let H be any nondecreasing function. Then the function h defined
by
h(a1, b1; · · · ; ak, bk) = E[H(Z1, · · · , Zk)
∣∣∣a1 ≤ Z1 ≤ b1, · · · , ak ≤ Zk ≤ bk
]is nondecreasing in all of its arguments.
37
Proof of Theorem 2. Suppose all other opponents bid according to the last minute bid-
ding strategy. Then, since nobody else bids, there is no information to update, and
hence, playing the repeated sealed-bid second price auction equilibrium, β1(vi) =
E[V(3)
∣∣∣V(1) = vi, V(2) = vi, h1,0
], according to your ex ante information is your best
response at the last minute t = T of stage 1, because if one deviates, that would trigger
other bidders to infer that that bidder has valuation vi ∈ [a, v], and by part 3 of lemma
1, would raise their last minute bid to
E[V(3)
∣∣∣V(m) = vj , V(m+1) = vj , Vj = vj , a < Vi < v]
≥ E[V(3)
∣∣∣V(m) = vj , V(m+1) = vj , Vj = vjv < Vi < v].
Proof of Theorem 3. Suppose there is another symmetric equilibrium with undomi-
nated strategies where all bidders play β1(vi;h1,t) containing bidding before the last
minute. By weak dominance of last minute bidding (Property 2 of Lemma 1), we
know that people will revise their bids at the last minute, and hence, do update their
information according to Bayes’ rule, using the last minute history h1,T .
For player i, consider deviating to solely using the last minute bidding strategy:
wait and bid
β1
(vi, h1,T
)= E
[V(3)
∣∣∣V(1) = vi, V(2) = vi, h1,T
]only at the last minute according to the then last minute history hm,T . The others’ last
minute bids are now
β1(vj ;h1,T ) = E[V(3)
∣∣∣V(1) = vi, V(2) = vj , h1,T
]Since the auction rule (Assumption 4c) specifies that later proxy bids must be
strictly higher than previous ones, and it is weakly dominated to bid above one’s valu-
ation, after seeing bidder j submit a proxy bid, bidder i’s beliefs are updated to be (at
least)
vj ≥ bj ≥ St ≥ 0.
38
Thus, as h1,t (strictly) increases, the conditional expectation
E[V(3)
∣∣∣V(1) = vi, V(2) = vi, h1,t
]is nondecreasing.
since player i has “retracted” all of her bids that were in h1,T except the last minute
one. Hence, we have
β1(vj ;h1,T ) = E[V(3)
∣∣∣V(1) = vi, V(2) = vj , h1,T
]≤ E
[V(3)
∣∣∣V(1) = vi, V(2) = vj , h1,T
]= β1(vj ; h1,T ).
Therefore, the expected payoffs for player i goes up (or stays constant) when she
deviates to the last minute bidding strategy. However, that means that β1(vj ;h1,T ) is
a (weakly) dominated strategy, and thus, should be ruled out by weak dominance.
Proof of Lemma 2. First note that for bidder i, bidding bk > vi is strictly dominated
by bidding bk = vi since the former adds a positive probability of winning and getting
a negative payoff (for paying bk > vi).
Furthermore, subgame perfection requires that after bidder i bid bk at time tk,
bidder j 6= i with valuation vj has to submit a counter bid bk+1 ≥ bk + s at time(tk + T
), as long as bk + s ≤ vj , if there were no other bids submitted during tk
and(tk + T
), or, equivalently, bidding history satisfies htk = htk+T . Hence, as the
standing bid increases, lesser bidders would be required to submit a counter bid since
the condition to counter bid, vj ≥ bn + s, is more likely to fail. Regardless of the
bidding path, as the standing bid increases, counter bids would cease to exist if and
only if the only bidder that might satisfy the condition vj ≥ bn + s is the the one who
submitted the last bid bn itself. In that case, no one will submit a counter bid during
the going-going-gone period(tn, tn + T
], and thus, the auction ends.
Since there is a minimum increment s, and the bidding sequence {bk} is strictly
increasing, there exist a finite n such that bn > v(2) − s, submitted by either the
highest bidder (with valuation v(1)) or the second-highest bidder (with valuation v(2)).
39
If it was the highest bidder submitting bn, the second-highest bidder would not counter
bid since bn + s > v(2).
However, as long as s < v(1) − v(2),50 the highest bidder can always counter any
possible bid bk of the second-highest bidder, which is at most v(2). Hence, the bidding
sequence is finite, and will end when someone bids bn in((v(2) − s), (v(2) + s)
].
This yields an efficient outcome, and as s → 0, the winning price bn converges to
v(2), the winning price in the (irrevocable) button auction.
Proof of Proposition 2. Facing opponents’ bidding strategy: Bid s, 2s, 3s, · · · , ks, · · ·
at the time T , 2T , · · · , kT , · · · is as if facing a raising price clock with the “tick” de-
fined by the minimal increment s. Hence, the question is whether one can “jump bid”
after a certain history to stop this price clock earlier than following the equilibrium
bidding strategy, which is to bid according to this price clock, since by jump bidding
to b, one can “block” out opponents who have valuations lower than b + s.
We now consider the case where at a current price ks, a bidder who has (k+2)s ≤
vi < (k + 3)s jump bidding to (k + 2)s at time (k + 1)T , instead of bidding (k + 1)s
as “scheduled.” (Note that if the current price is below ks, the “loss” is even larger for
jump bidding, while the “gains” of squeezing out opponents remain the same. Thus,
the case considered here provides the upper bound of gains from deviation.) The gains
are from blocking out y1 ∈ [v, (k + 3)s) since (k + 2)s + s = (k + 3)s. The losses
occur when y1 ∈ [ks, (k + 1)s], and hence, one could have won and paid (k + 1)s
instead. Therefore, in equilibrium, we require
∆u = Pr(v < y1 < (k + 3)s
)·[v − (k + 2)s
]+ Pr
(ks < y1 < (k + 1)s
)·[v − (k + 2)s− v + (k + 1)s
]=
[F2
((k + 3)s
)− F2(v)
]· [v − (k + 2)s] +
[F2
((k + 1)s
)− F2(ks)
]· (−s)
≤ 0
50If s ≥ v(1)−v(2), we still get approximate efficiency since the bidder with v(2) wins the item and further
gains from trade is v(1) − v(2), which is at most s.
40
Or,
F2((k+3)s)−F2(v)(k+3)s−v
F2((k+1)s)−F2(ks)s
≤ s
(k + 3)s− v· s
v − (k + 2)s
=1[
(k + 3)− vs
]·[
vs − (k + 2)
]=
114 −
[vs − (k + 5
2)]2
≤ 4 forv
s∈ [k + 2, k + 3].
Since F ′2 = f2 exists for all v(2), by the mean value theorem, there exists ξ1 ∈ [v, (k +
3)s] and ξ2 ∈ [ks, (k + 1)s] such that
f2(ξ1)f2(ξ2)
=F2((k+3)s)−F2(v)
(k+3)s−v
F2((k+1)s)−F2(ks)s
≤ 4
because by assumption,
f2(y + η)f2(y)
≤ 4 for all y ∈ [v, v] and η ∈ [s, 3s].
Thus, the requirement of non-deviation is satisfied, and the proposed symmetric strat-
egy consists an equilibrium.
As s → 0, the smoothness assumption requires smaller and smaller ranges of η,
and as η → 0,f2(y + η)
f2(y)→ 1 < 4
trivially.
Finally, when s → 0, the outcome is bidders jointly pushing up the standing price
by s every T . Letting T = o(s) (so that the clock stops in finite time), we have
constructed a “price clock” which is equivalent to what we see in a button auction,
and the cease-to-bid criteria of vi ≥ ks = bi(ks) is equivalent to the dropping out
strategy.
41
References
Ariely, Dan, Axel Ockenfels, and Alvin E. Roth, “An Experimental Analysis of
Ending Rules in Internet Auctions,” Rand Journal of Economics, 2005, 36 (4), 891–
908.
Ashenfelter, Orley, “How Auctions Work for Wine and Art,” Journal of Economic
Perspectives, 1989, 3 (3), 23–36.
Avery, Christopher, “Strategic Jump Bidding in English Auctions,” Review of Eco-
nomic Studies, 1998, 65 (2), 185–210.
Bajari, Patrick and Ali Hortacsu, “The Winner’s Curse, Reserve Rrices, and En-
dogenous Entry: Empirical Insights from eBay Auctions,” Rand Journal of Eco-
nomics, 2003, 34 (2), 329–355.
Binmore, Kenneth G. and Paul Klemperer, “The Biggest Auction Ever: the Sale of
the British 3G Telecom Licences,” Economic Journal, 2002, 112 (478), C74–C96.
Izmalkov, Sergei, “English Auctions with Reentry,” 2003. unpublished paper.
McAfee, R. Preston and Daniel Vincent, “The Declining Price Anomaly,” Journal
of Economic Theory, 1993, 60 (1), 191–212.
Milgrom, Paul R. and Robert J. Weber, “A Theory of Auctions and Competitive
Bidding,” Econometrica, 1982, 50 (5), 1089–1122.
and , “A Theory of Auctions and Competitive Bidding, II,” in Paul Klemperer,
ed., The Economic Theory of Auctions, Vol. 2, Edward Elgar Publishing, 2000,
pp. 179–194.
Ockenfels, Axel and Alvin E. Roth, “Late and Multiple Bidding in Second Price
Internet Auctions: Theory and Evidence Concerning Different Rules for Ending an
Auction,” Games and Economic Behavior, 2006, 55 (2), 297320.
42
Rasmusen, Eric, “Getting Carried Away in Auctions as Imperfect Value Discovery,”
2005. unpublished paper.
Rasmusen, Eric Bennett, “Strategic Implications of Uncertainty Over One’s Own
Private Value in Auctions,” Advances in Theoretical Economics, 2006, 6 (1).
Roth, Alvin E. and Axel Ockenfels, “Last-Minute Bidding and the Rules for Ending
Second-Price Auctions: Evidence from eBay and Amazon Auctions on the Inter-
net,” American Economic Review, 2002, 92 (4), 1093–1103.
Schindler, Julia, “Late Bidding on the Internet,” 2003. unpublished paper.
Vickrey, William, “Counterspeculation, Auctions, and Competitive Sealed Tenders,”
Journal of Finance, 1961, 16 (1), 8–37.
Wang, Joseph Taoyi, “Equilibrium in Sequential Second Price Auctions,” 2006. un-
published paper.
Weber, Robert J., “Multiple-Object Auctions,” in Richard Engelbrecht-Wiggans,
Martin Shubik, and Robert M. Stark, eds., Auctions, Bidding, and Contracting:
Uses and Theory, New York: New York University Press, 1983, pp. 165–191.
Wilson, Robert, “A Comment on: David Hughart, Informational Asymmetry, Bidding
Strategies, and the Marketing of Offshore Petroleum Leases,” 1975. unpublished
paper.
43